Designer-Grade Image Prompts for a Few Cents: Pair Gemini 4 Argon With Nano Banana 2.1

Image models are only as good as the prompt, and most of our prompts need improvement: a subject, maybe a style, nothing about composition, light, palette or where the text goes. A language model that follows instructions can write the better prompt from a plain sentence for a fraction of a cent. So the pipeline is two calls: Argon writes, Nano Banana paints. The same two steps run as two API calls you can copy into any app, or as one chat message through the GMI MCP server, where the assistant writes the prompt and calls the image model itself.

The two models

Gemini 4 Argon is the writer. Google announced it on September 30, gave the first access to vetted cyber defenders through its Fairwind Program, because the model can find, validate and patch software vulnerabilities on its own, and opened the API on October 13. Three traits make it the right model to write prompts. It follows instructions: Google's launch post puts it first on AutomationBench, Zapier's test of whether a model completes a business task end to end rather than describing it, at 51.3%, and the chart below shows the same post's numbers for long-horizon coding and finance agent tasks against Claude Opus 5.5 and GPT-6 Astra. It reads images, which is what lets it look at a photo before writing an edit instruction. And its output limit is a million tokens, up from 64K on the previous generation: irrelevant for a two-hundred-word prompt, but a sign of where Google is pointing it. Google also calls it its most resilient model yet against indirect prompt injection, which matters once a prompt-writer reads text other people wrote.

Google's Gemini 4 Argon announcement, September 30, 2026

Nano Banana 2.1 is the painter. It is Google's October 6 update to Nano Banana 2, the Flash-tier image model, so it keeps Flash-level speed, and the release notes read like a list of fixed complaints: better text rendering, subject consistency across up to fourteen reference images (up to four characters and ten objects), and wide and tall frames all the way to 1:8 and 8:1 at 1K, 2K and 4K. It thinks before it draws, and on GMI you choose how much. In GMI's own image benchmark from earlier this month it scored 4.8 out of 5 across five prompts, about the same as the top of the set, with the most natural skin in the portrait test and an exact sign in the signage test, and at 1K it returned every image in 14 to 15 seconds.

GMI image benchmark, 13 models, 5 prompts, October 2026

What a designer-grade prompt looks like

Here is the second job in full, so you can see what Argon adds to a sentence. My brief:

An October wallpaper for my phone. Quiet, something I won't get sick of, dark enough that the clock stays readable.

Argon's prompt, with 9:16 and 2K chosen for the frame:

A serene, minimalist October phone wallpaper. In the lower third, a single slender branch of a Japanese maple with a few deep burnt-orange and muted crimson leaves extends gently from the right edge, speckled with tiny raindrops. The upper two-thirds of the frame is completely open, uncluttered negative space featuring a smooth, dark slate-blue and charcoal twilight sky with soft rolling mist in the far distance, keeping the top area dark and quiet so a phone clock is easily readable. Moody low-key dusk lighting, shallow depth of field, shot on a 50mm lens at f/2.0, subtle fine film grain, palette of charcoal, deep navy, muted rust, and umber, calm contemplative atmosphere, no text.

Avoid: bright sky, harsh sunlight, busy background, text, letters, watermark, neon colors, jack-o-lanterns, halloween cliches, clutter in the top half

Six things in that prompt are what a designer would have put in the brief and I did not. It puts the branch in the lower third and keeps the upper two thirds open, because a wallpaper has a clock on it. It names the lens and the aperture, 50mm at f/2.0, which is how you ask an image model for a particular depth of field without saying the words. It names the palette: charcoal, deep navy, muted rust, umber. It sets the light, low-key dusk. It adds the small details that make a photo read as a photo, raindrops on the leaves and fine film grain. And the negative list rules out the Halloween cliches that the word October invites. That is 257 output tokens, a third of a cent, written in seven seconds.

The other four prompts follow the same pattern: an 85mm lens at f/2.8 for the bakery ad, "isometric 3D" and "editorial tech illustration" for the blog cover, and in every job a sentence about where text goes or where it must not. None of that was in my sentences.

Five jobs, one sentence each

Every job below started as the one sentence in bold. Argon wrote the prompt, picked the aspect ratio and the resolution, and listed what to avoid. Nano Banana 2.1 painted it. The images are the first take; nothing was regenerated or retouched.

"Cover for my indie tech zine called SIGNAL, issue 14, October 2026. A dark desk at night, one lamp. Cover lines: The 2 a.m. commit, Why small teams ship faster, Inside the GPU shortage." Argon chose 3:4 at 2K, put the masthead, the issue line and every cover line in quotes, added a CRT showing green code and a brass lamp, and asked for dark negative space at the top and down the left margin "for typography". Every word is spelled as typed. One cover line came out amber rather than the off-white the prompt asked for, which is the only miss in the five jobs.

"An October wallpaper for my phone. Quiet, something I won't get sick of, dark enough that the clock stays readable." Argon chose 9:16 at 2K, put a single maple branch in the lower third, and wrote "the upper two-thirds of the frame is completely open" so the clock has room. My sentence only said the clock should stay readable; the open two-thirds was Argon's reading of that.

"Instagram ad for my neighborhood bakery's new sourdough. $9, this Saturday only." Argon chose 4:5 at 2K, wrote the copy as "FRESH SOURDOUGH" and "THIS SATURDAY ONLY" in quotes, added a gold price badge, and listed "misspelled text" and "gibberish" among the things to avoid. Three pieces of text, all correct.

"Cover image for a blog post about long-context language models reading an entire codebase at once. No text in the image." Argon chose 16:9 at 1K and, without being asked, kept the left third as clean negative space "suitable for a blog title overlay", because covers get text laid over them.

The photo edit. I handed Argon the finished cover with one sentence: masthead to white, the three cover lines to lime green, everything else identical. Argon looked at the photo and wrote an instruction that names what it saw: the "SIGNAL" masthead, the three cover lines by their words, and the lamp, mug, keyboard and CRT monitor that must not move. The photo goes to Argon inline as a base64 data URL, and to Nano Banana as the public URL of the uploaded file in the image field. The result is the same scene with the two colors changed and nothing else moved, and it fixed the amber cover line along the way. That is what makes the pair usable for revisions, not only first drafts.

Two ways to run it

One chat message. In Claude with the GMI connector, this is the whole job:

Write a detailed image prompt for a magazine cover for my indie tech zine called SIGNAL, dark desk at night, cover lines "The 2 a.m. commit", "Why small teams ship faster", "Inside the GPU shortage". Then generate it with gemini-nano-banana-2.1 at 3:4, 1K, and show me the image and what it cost.

The assistant writes the prompt, calls estimate_generation, quotes the price, calls submit_generation, and the image comes back inline. For an edit, attach the photo and ask for an edit instruction instead; the connector uploads it and passes the URL as the reference image.

Two API calls. The first goes to Argon on the chat-completions endpoint and asks for a JSON object:

curl https://api.gmi-serving.com/v1/chat/completions \
  -H "Authorization: Bearer $GMI_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-4-argon",
    "messages": [
      {"role": "system", "content": "You write image prompts. Reply with one JSON object: {\"prompt\", \"aspect_ratio\", \"image_size\", \"negative\"}. Put any text that must appear in the image in quotes, spelled exactly."},
      {"role": "user", "content": "Phone wallpaper for October. Quiet, dark enough that the clock stays readable."}
    ]
  }'

The second posts Argon's answer to the request queue as Nano Banana 2.1, then polls until status is success and reads outcome.media_urls:

curl https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests \
  -H "Authorization: Bearer $GMI_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-nano-banana-2.1",
    "payload": {
      "prompt": "<prompt from Argon>",
      "aspect_ratio": "9:16",
      "image_size": "2K",
      "thinking": "medium",
      "image_output_format": "png"
    }
  }'

curl https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests/<request_id> \
  -H "Authorization: Bearer $GMI_API_KEY"

For a photo edit, request an upload URL from .../requestqueue/apikey/upload-url with {"file_type": "png"}, PUT the file to it, and pass the returned public_url as image in the payload. The response carries two URLs per request: at 1K they are the same file, at 2K and 4K the first is a 1024-pixel preview and the second is the full-resolution image. One image is generated and billed.

What it cost

Five images and five prompts came to about thirty cents at list price. Argon's share is small: the text-only prompt calls were about 320 tokens in and 240 to 300 out, a third of a cent each, and the edit call, which carries the photo, was 1,403 tokens in and two-thirds of a cent. The cover call's usage was not logged, so its line below is estimated from the other four.

Resolution is the only lever that moves the bill: 1K to 4K triples the image price, and the thinking level is billed as text output, and on prompts this short it made no visible difference in GMI's benchmark. Every number here is a list price from the GMI Cloud inference engine on a Tier 5 account; cost and wait times vary with account tier, queue load and provider capacity, and a dedicated endpoint runs on reserved capacity without the queue.

Where to be careful

The benchmark figures are Google's. The AutomationBench, DeepSWE and Vals Finance numbers come from Google's launch post, and no third party had reproduced them when I wrote this. They are why I tried Argon as the writer, not proof that it is the best one. The image side rests on GMI's own five-prompt benchmark with a single rater, so a tenth of a point there is not a ranking.

Send Argon its images inline. On GMI, Gemini models read a photo as a base64 data URL inside image_url; they do not fetch HTTP links. Nano Banana is the opposite: it takes the public URL of an uploaded file. The edit job above uses both paths.

The writer is a general model at $2 and $10 per million tokens. A cheaper model can write a usable prompt, and for a plain text-to-image job it may be enough. Argon earned its place on the edit, where it had to read the finished cover and name what must not move, and on the briefs where it reserved negative space nobody asked for.

Check the text anyway. The one miss across five jobs was a color, not a word, but image text is still the thing to look at before you ship. Argon puts every string that must appear in the image in quotes, spelled exactly, and Nano Banana rendered all of them as typed in this run. Keep that quote rule in your system prompt.

Try It on GMI Cloud

One call to the writer. Swap the brief for your own and keep the system prompt:

curl https://api.gmi-serving.com/v1/chat/completions \
  -H "Authorization: Bearer $GMI_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-4-argon",
    "messages": [
      {"role": "system", "content": "You write image prompts for Nano Banana 2.1. Reply with one JSON object: {\"prompt\", \"aspect_ratio\", \"image_size\", \"negative\"}. Be concrete about subject, composition, lighting and palette. Put any text that must appear in the image in quotes, spelled exactly."},
      {"role": "user", "content": "A cover image for a blog post about long-context models. No text in the image."}
    ]
  }'

Then paste the prompt, aspect_ratio and image_size from the answer into the Nano Banana request above. The whole pair costs under five cents at 1K.

The Question to Answer

The prompt was the hard part, and a two-dollar-per-million language model writes a better one than I do in about ten seconds. So before you pick an image model for a feature, decide who writes the prompt.

Ready to build? Get Started